* [PATCH 1/6] DIO: Separate fields only used in the submission path from struct dio
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
@ 2011-06-22 19:18 ` Andi Kleen
2011-06-22 19:18 ` [PATCH 2/6] DIO: Fix a wrong comment Andi Kleen
` (5 subsequent siblings)
6 siblings, 0 replies; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 19:18 UTC (permalink / raw)
To: linux-fsdevel; +Cc: linux-kernel, dehrenberg, Andi Kleen
From: Andi Kleen <ak@linux.intel.com>
This large, but largely mechanic, patch moves all fields in struct dio
that are only used in the submission path into a separate on stack
data structure. This has the advantage that the memory is very likely
cache hot, which is not guaranteed for memory fresh out of kmalloc.
This also gives gcc more optimization potential because it can easier
determine that there are no external aliases for these variables.
The sdio initialization is a initialization now instead of memset.
This allows gcc to break sdio into individual fields and optimize
away unnecessary zeroing (after all the functions are inlined)
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
fs/direct-io.c | 395 +++++++++++++++++++++++++++++---------------------------
1 files changed, 203 insertions(+), 192 deletions(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index ac5f164..e1360c6 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -55,13 +55,10 @@
* blocksize.
*/
-struct dio {
- /* BIO submission state */
+/* dio_state only used in the submission path */
+
+struct dio_submit {
struct bio *bio; /* bio under assembly */
- struct inode *inode;
- int rw;
- loff_t i_size; /* i_size when submitted */
- int flags; /* doesn't change */
unsigned blkbits; /* doesn't change */
unsigned blkfactor; /* When we're using an alignment which
is finer than the filesystem's soft
@@ -71,7 +68,7 @@ struct dio {
unsigned start_zero_done; /* flag: sub-blocksize zeroing has
been performed at the start of a
write */
- int pages_in_io; /* approximate total IO pages */
+ int pages_in_io; /* approximate total IO pages */
size_t size; /* total request size (doesn't change)*/
sector_t block_in_file; /* Current offset into the underlying
file in dio_block units. */
@@ -81,14 +78,13 @@ struct dio {
int boundary; /* prev block is at a boundary */
int reap_counter; /* rate limit reaping */
get_block_t *get_block; /* block mapping function */
- dio_iodone_t *end_io; /* IO completion function */
- dio_submit_t *submit_io; /* IO submition function */
+ dio_submit_t *submit_io; /* IO submition function */
+
loff_t logical_offset_in_bio; /* current first logical block in bio */
sector_t final_block_in_bio; /* current final block in bio + 1 */
sector_t next_block_for_io; /* next block to be put under IO,
in dio_blocks units */
- struct buffer_head map_bh; /* last get_block() result */
-
+
/*
* Deferred addition of a page to the dio. These variables are
* private to dio_send_cur_page(), submit_page_section() and
@@ -100,18 +96,6 @@ struct dio {
sector_t cur_page_block; /* Where it starts */
loff_t cur_page_fs_offset; /* Offset in file */
- /* BIO completion state */
- spinlock_t bio_lock; /* protects BIO fields below */
- unsigned long refcount; /* direct_io_worker() and bios */
- struct bio *bio_list; /* singly linked via bi_private */
- struct task_struct *waiter; /* waiting task (NULL if none) */
-
- /* AIO related stuff */
- struct kiocb *iocb; /* kiocb */
- int is_async; /* is IO async ? */
- int io_error; /* IO error in completion path */
- ssize_t result; /* IO result */
-
/*
* Page fetching state. These variables belong to dio_refill_pages().
*/
@@ -125,6 +109,30 @@ struct dio {
*/
unsigned head; /* next page to process */
unsigned tail; /* last valid page + 1 */
+};
+
+/* dio_state communicated between submission path and end_io */
+struct dio {
+ int flags; /* doesn't change */
+ struct inode *inode;
+ int rw;
+ loff_t i_size; /* i_size when submitted */
+ dio_iodone_t *end_io; /* IO completion function */
+ struct buffer_head map_bh; /* last get_block() result */
+
+
+ /* BIO completion state */
+ spinlock_t bio_lock; /* protects BIO fields below */
+ unsigned long refcount; /* direct_io_worker() and bios */
+ struct bio *bio_list; /* singly linked via bi_private */
+ struct task_struct *waiter; /* waiting task (NULL if none) */
+
+ /* AIO related stuff */
+ struct kiocb *iocb; /* kiocb */
+ int is_async; /* is IO async ? */
+ int io_error; /* IO error in completion path */
+ ssize_t result; /* IO result */
+
int page_errors; /* errno from get_user_pages() */
/*
@@ -132,33 +140,33 @@ struct dio {
* allocation time. Don't add new fields after pages[] unless you
* wish that they not be zeroed.
*/
- struct page *pages[DIO_PAGES]; /* page buffer */
+ struct page *pages[DIO_PAGES]; /* page buffer */
};
/*
* How many pages are in the queue?
*/
-static inline unsigned dio_pages_present(struct dio *dio)
+static inline unsigned dio_pages_present(struct dio *dio, struct dio_submit *sdio)
{
- return dio->tail - dio->head;
+ return sdio->tail - sdio->head;
}
/*
* Go grab and pin some userspace pages. Typically we'll get 64 at a time.
*/
-static int dio_refill_pages(struct dio *dio)
+static int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
{
int ret;
int nr_pages;
- nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
+ nr_pages = min(sdio->total_pages - sdio->curr_page, DIO_PAGES);
ret = get_user_pages_fast(
- dio->curr_user_address, /* Where from? */
+ sdio->curr_user_address, /* Where from? */
nr_pages, /* How many pages? */
dio->rw == READ, /* Write to memory? */
&dio->pages[0]); /* Put results here */
- if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
+ if (ret < 0 && sdio->blocks_available && (dio->rw & WRITE)) {
struct page *page = ZERO_PAGE(0);
/*
* A memory fault, but the filesystem has some outstanding
@@ -169,17 +177,17 @@ static int dio_refill_pages(struct dio *dio)
dio->page_errors = ret;
page_cache_get(page);
dio->pages[0] = page;
- dio->head = 0;
- dio->tail = 1;
+ sdio->head = 0;
+ sdio->tail = 1;
ret = 0;
goto out;
}
if (ret >= 0) {
- dio->curr_user_address += ret * PAGE_SIZE;
- dio->curr_page += ret;
- dio->head = 0;
- dio->tail = ret;
+ sdio->curr_user_address += ret * PAGE_SIZE;
+ sdio->curr_page += ret;
+ sdio->head = 0;
+ sdio->tail = ret;
ret = 0;
}
out:
@@ -192,17 +200,17 @@ out:
* decent number of pages, less frequently. To provide nicer use of the
* L1 cache.
*/
-static struct page *dio_get_page(struct dio *dio)
+static struct page *dio_get_page(struct dio *dio, struct dio_submit *sdio)
{
- if (dio_pages_present(dio) == 0) {
+ if (dio_pages_present(dio, sdio) == 0) {
int ret;
- ret = dio_refill_pages(dio);
+ ret = dio_refill_pages(dio, sdio);
if (ret)
return ERR_PTR(ret);
- BUG_ON(dio_pages_present(dio) == 0);
+ BUG_ON(dio_pages_present(dio, sdio) == 0);
}
- return dio->pages[dio->head++];
+ return dio->pages[sdio->head++];
}
/**
@@ -326,8 +334,9 @@ void dio_end_io(struct bio *bio, int error)
EXPORT_SYMBOL_GPL(dio_end_io);
static void
-dio_bio_alloc(struct dio *dio, struct block_device *bdev,
- sector_t first_sector, int nr_vecs)
+dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
+ struct block_device *bdev,
+ sector_t first_sector, int nr_vecs)
{
struct bio *bio;
@@ -344,8 +353,8 @@ dio_bio_alloc(struct dio *dio, struct block_device *bdev,
else
bio->bi_end_io = dio_bio_end_io;
- dio->bio = bio;
- dio->logical_offset_in_bio = dio->cur_page_fs_offset;
+ sdio->bio = bio;
+ sdio->logical_offset_in_bio = sdio->cur_page_fs_offset;
}
/*
@@ -355,9 +364,9 @@ dio_bio_alloc(struct dio *dio, struct block_device *bdev,
*
* bios hold a dio reference between submit_bio and ->end_io.
*/
-static void dio_bio_submit(struct dio *dio)
+static void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
{
- struct bio *bio = dio->bio;
+ struct bio *bio = sdio->bio;
unsigned long flags;
bio->bi_private = dio;
@@ -369,24 +378,24 @@ static void dio_bio_submit(struct dio *dio)
if (dio->is_async && dio->rw == READ)
bio_set_pages_dirty(bio);
- if (dio->submit_io)
- dio->submit_io(dio->rw, bio, dio->inode,
- dio->logical_offset_in_bio);
+ if (sdio->submit_io)
+ sdio->submit_io(dio->rw, bio, dio->inode,
+ sdio->logical_offset_in_bio);
else
submit_bio(dio->rw, bio);
- dio->bio = NULL;
- dio->boundary = 0;
- dio->logical_offset_in_bio = 0;
+ sdio->bio = NULL;
+ sdio->boundary = 0;
+ sdio->logical_offset_in_bio = 0;
}
/*
* Release any resources in case of a failure
*/
-static void dio_cleanup(struct dio *dio)
+static void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
{
- while (dio_pages_present(dio))
- page_cache_release(dio_get_page(dio));
+ while (dio_pages_present(dio, sdio))
+ page_cache_release(dio_get_page(dio, sdio));
}
/*
@@ -476,11 +485,11 @@ static void dio_await_completion(struct dio *dio)
*
* This also helps to limit the peak amount of pinned userspace memory.
*/
-static int dio_bio_reap(struct dio *dio)
+static int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
{
int ret = 0;
- if (dio->reap_counter++ >= 64) {
+ if (sdio->reap_counter++ >= 64) {
while (dio->bio_list) {
unsigned long flags;
struct bio *bio;
@@ -494,14 +503,14 @@ static int dio_bio_reap(struct dio *dio)
if (ret == 0)
ret = ret2;
}
- dio->reap_counter = 0;
+ sdio->reap_counter = 0;
}
return ret;
}
/*
* Call into the fs to map some more disk blocks. We record the current number
- * of available blocks at dio->blocks_available. These are in units of the
+ * of available blocks at sdio->blocks_available. These are in units of the
* fs blocksize, (1 << inode->i_blkbits).
*
* The fs is allowed to map lots of blocks at once. If it wants to do that,
@@ -522,7 +531,7 @@ static int dio_bio_reap(struct dio *dio)
* buffer_mapped(). However the direct-io code will only process holes one
* block at a time - it will repeatedly call get_block() as it walks the hole.
*/
-static int get_more_blocks(struct dio *dio)
+static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
{
int ret;
struct buffer_head *map_bh = &dio->map_bh;
@@ -538,11 +547,11 @@ static int get_more_blocks(struct dio *dio)
*/
ret = dio->page_errors;
if (ret == 0) {
- BUG_ON(dio->block_in_file >= dio->final_block_in_request);
- fs_startblk = dio->block_in_file >> dio->blkfactor;
- dio_count = dio->final_block_in_request - dio->block_in_file;
- fs_count = dio_count >> dio->blkfactor;
- blkmask = (1 << dio->blkfactor) - 1;
+ BUG_ON(sdio->block_in_file >= sdio->final_block_in_request);
+ fs_startblk = sdio->block_in_file >> sdio->blkfactor;
+ dio_count = sdio->final_block_in_request - sdio->block_in_file;
+ fs_count = dio_count >> sdio->blkfactor;
+ blkmask = (1 << sdio->blkfactor) - 1;
if (dio_count & blkmask)
fs_count++;
@@ -562,12 +571,12 @@ static int get_more_blocks(struct dio *dio)
*/
create = dio->rw & WRITE;
if (dio->flags & DIO_SKIP_HOLES) {
- if (dio->block_in_file < (i_size_read(dio->inode) >>
- dio->blkbits))
+ if (sdio->block_in_file < (i_size_read(dio->inode) >>
+ sdio->blkbits))
create = 0;
}
- ret = (*dio->get_block)(dio->inode, fs_startblk,
+ ret = (*sdio->get_block)(dio->inode, fs_startblk,
map_bh, create);
}
return ret;
@@ -576,20 +585,21 @@ static int get_more_blocks(struct dio *dio)
/*
* There is no bio. Make one now.
*/
-static int dio_new_bio(struct dio *dio, sector_t start_sector)
+static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
+ sector_t start_sector)
{
sector_t sector;
int ret, nr_pages;
- ret = dio_bio_reap(dio);
+ ret = dio_bio_reap(dio, sdio);
if (ret)
goto out;
- sector = start_sector << (dio->blkbits - 9);
- nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
+ sector = start_sector << (sdio->blkbits - 9);
+ nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
nr_pages = min(nr_pages, BIO_MAX_PAGES);
BUG_ON(nr_pages <= 0);
- dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
- dio->boundary = 0;
+ dio_bio_alloc(dio, sdio, dio->map_bh.b_bdev, sector, nr_pages);
+ sdio->boundary = 0;
out:
return ret;
}
@@ -601,21 +611,21 @@ out:
*
* Return zero on success. Non-zero means the caller needs to start a new BIO.
*/
-static int dio_bio_add_page(struct dio *dio)
+static int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
{
int ret;
- ret = bio_add_page(dio->bio, dio->cur_page,
- dio->cur_page_len, dio->cur_page_offset);
- if (ret == dio->cur_page_len) {
+ ret = bio_add_page(sdio->bio, sdio->cur_page,
+ sdio->cur_page_len, sdio->cur_page_offset);
+ if (ret == sdio->cur_page_len) {
/*
* Decrement count only, if we are done with this page
*/
- if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
- dio->pages_in_io--;
- page_cache_get(dio->cur_page);
- dio->final_block_in_bio = dio->cur_page_block +
- (dio->cur_page_len >> dio->blkbits);
+ if ((sdio->cur_page_len + sdio->cur_page_offset) == PAGE_SIZE)
+ sdio->pages_in_io--;
+ page_cache_get(sdio->cur_page);
+ sdio->final_block_in_bio = sdio->cur_page_block +
+ (sdio->cur_page_len >> sdio->blkbits);
ret = 0;
} else {
ret = 1;
@@ -633,14 +643,14 @@ static int dio_bio_add_page(struct dio *dio)
* The caller of this function is responsible for removing cur_page from the
* dio, and for dropping the refcount which came from that presence.
*/
-static int dio_send_cur_page(struct dio *dio)
+static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio)
{
int ret = 0;
- if (dio->bio) {
- loff_t cur_offset = dio->cur_page_fs_offset;
- loff_t bio_next_offset = dio->logical_offset_in_bio +
- dio->bio->bi_size;
+ if (sdio->bio) {
+ loff_t cur_offset = sdio->cur_page_fs_offset;
+ loff_t bio_next_offset = sdio->logical_offset_in_bio +
+ sdio->bio->bi_size;
/*
* See whether this new request is contiguous with the old.
@@ -656,28 +666,28 @@ static int dio_send_cur_page(struct dio *dio)
* be the next logical offset in the bio, submit the bio we
* have.
*/
- if (dio->final_block_in_bio != dio->cur_page_block ||
+ if (sdio->final_block_in_bio != sdio->cur_page_block ||
cur_offset != bio_next_offset)
- dio_bio_submit(dio);
+ dio_bio_submit(dio, sdio);
/*
* Submit now if the underlying fs is about to perform a
* metadata read
*/
- else if (dio->boundary)
- dio_bio_submit(dio);
+ else if (sdio->boundary)
+ dio_bio_submit(dio, sdio);
}
- if (dio->bio == NULL) {
- ret = dio_new_bio(dio, dio->cur_page_block);
+ if (sdio->bio == NULL) {
+ ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
if (ret)
goto out;
}
- if (dio_bio_add_page(dio) != 0) {
- dio_bio_submit(dio);
- ret = dio_new_bio(dio, dio->cur_page_block);
+ if (dio_bio_add_page(dio, sdio) != 0) {
+ dio_bio_submit(dio, sdio);
+ ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
if (ret == 0) {
- ret = dio_bio_add_page(dio);
+ ret = dio_bio_add_page(dio, sdio);
BUG_ON(ret != 0);
}
}
@@ -703,7 +713,7 @@ out:
* page to the dio instead.
*/
static int
-submit_page_section(struct dio *dio, struct page *page,
+submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
unsigned offset, unsigned len, sector_t blocknr)
{
int ret = 0;
@@ -718,20 +728,20 @@ submit_page_section(struct dio *dio, struct page *page,
/*
* Can we just grow the current page's presence in the dio?
*/
- if ( (dio->cur_page == page) &&
- (dio->cur_page_offset + dio->cur_page_len == offset) &&
- (dio->cur_page_block +
- (dio->cur_page_len >> dio->blkbits) == blocknr)) {
- dio->cur_page_len += len;
+ if ( (sdio->cur_page == page) &&
+ (sdio->cur_page_offset + sdio->cur_page_len == offset) &&
+ (sdio->cur_page_block +
+ (sdio->cur_page_len >> sdio->blkbits) == blocknr)) {
+ sdio->cur_page_len += len;
/*
- * If dio->boundary then we want to schedule the IO now to
+ * If sdio->boundary then we want to schedule the IO now to
* avoid metadata seeks.
*/
- if (dio->boundary) {
- ret = dio_send_cur_page(dio);
- page_cache_release(dio->cur_page);
- dio->cur_page = NULL;
+ if (sdio->boundary) {
+ ret = dio_send_cur_page(dio, sdio);
+ page_cache_release(sdio->cur_page);
+ sdio->cur_page = NULL;
}
goto out;
}
@@ -739,20 +749,20 @@ submit_page_section(struct dio *dio, struct page *page,
/*
* If there's a deferred page already there then send it.
*/
- if (dio->cur_page) {
- ret = dio_send_cur_page(dio);
- page_cache_release(dio->cur_page);
- dio->cur_page = NULL;
+ if (sdio->cur_page) {
+ ret = dio_send_cur_page(dio, sdio);
+ page_cache_release(sdio->cur_page);
+ sdio->cur_page = NULL;
if (ret)
goto out;
}
page_cache_get(page); /* It is in dio */
- dio->cur_page = page;
- dio->cur_page_offset = offset;
- dio->cur_page_len = len;
- dio->cur_page_block = blocknr;
- dio->cur_page_fs_offset = dio->block_in_file << dio->blkbits;
+ sdio->cur_page = page;
+ sdio->cur_page_offset = offset;
+ sdio->cur_page_len = len;
+ sdio->cur_page_block = blocknr;
+ sdio->cur_page_fs_offset = sdio->block_in_file << sdio->blkbits;
out:
return ret;
}
@@ -784,19 +794,19 @@ static void clean_blockdev_aliases(struct dio *dio)
* `end' is zero if we're doing the start of the IO, 1 at the end of the
* IO.
*/
-static void dio_zero_block(struct dio *dio, int end)
+static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
{
unsigned dio_blocks_per_fs_block;
unsigned this_chunk_blocks; /* In dio_blocks */
unsigned this_chunk_bytes;
struct page *page;
- dio->start_zero_done = 1;
- if (!dio->blkfactor || !buffer_new(&dio->map_bh))
+ sdio->start_zero_done = 1;
+ if (!sdio->blkfactor || !buffer_new(&dio->map_bh))
return;
- dio_blocks_per_fs_block = 1 << dio->blkfactor;
- this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
+ dio_blocks_per_fs_block = 1 << sdio->blkfactor;
+ this_chunk_blocks = sdio->block_in_file & (dio_blocks_per_fs_block - 1);
if (!this_chunk_blocks)
return;
@@ -808,14 +818,14 @@ static void dio_zero_block(struct dio *dio, int end)
if (end)
this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
- this_chunk_bytes = this_chunk_blocks << dio->blkbits;
+ this_chunk_bytes = this_chunk_blocks << sdio->blkbits;
page = ZERO_PAGE(0);
- if (submit_page_section(dio, page, 0, this_chunk_bytes,
- dio->next_block_for_io))
+ if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
+ sdio->next_block_for_io))
return;
- dio->next_block_for_io += this_chunk_blocks;
+ sdio->next_block_for_io += this_chunk_blocks;
}
/*
@@ -834,9 +844,9 @@ static void dio_zero_block(struct dio *dio, int end)
* it should set b_size to PAGE_SIZE or more inside get_block(). This gives
* fine alignment but still allows this function to work in PAGE_SIZE units.
*/
-static int do_direct_IO(struct dio *dio)
+static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
{
- const unsigned blkbits = dio->blkbits;
+ const unsigned blkbits = sdio->blkbits;
const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
struct page *page;
unsigned block_in_page;
@@ -844,10 +854,10 @@ static int do_direct_IO(struct dio *dio)
int ret = 0;
/* The I/O can start at any block offset within the first page */
- block_in_page = dio->first_block_in_page;
+ block_in_page = sdio->first_block_in_page;
- while (dio->block_in_file < dio->final_block_in_request) {
- page = dio_get_page(dio);
+ while (sdio->block_in_file < sdio->final_block_in_request) {
+ page = dio_get_page(dio, sdio);
if (IS_ERR(page)) {
ret = PTR_ERR(page);
goto out;
@@ -859,14 +869,14 @@ static int do_direct_IO(struct dio *dio)
unsigned this_chunk_blocks; /* # of blocks */
unsigned u;
- if (dio->blocks_available == 0) {
+ if (sdio->blocks_available == 0) {
/*
* Need to go and map some more disk
*/
unsigned long blkmask;
unsigned long dio_remainder;
- ret = get_more_blocks(dio);
+ ret = get_more_blocks(dio, sdio);
if (ret) {
page_cache_release(page);
goto out;
@@ -874,18 +884,18 @@ static int do_direct_IO(struct dio *dio)
if (!buffer_mapped(map_bh))
goto do_holes;
- dio->blocks_available =
- map_bh->b_size >> dio->blkbits;
- dio->next_block_for_io =
- map_bh->b_blocknr << dio->blkfactor;
+ sdio->blocks_available =
+ map_bh->b_size >> sdio->blkbits;
+ sdio->next_block_for_io =
+ map_bh->b_blocknr << sdio->blkfactor;
if (buffer_new(map_bh))
clean_blockdev_aliases(dio);
- if (!dio->blkfactor)
+ if (!sdio->blkfactor)
goto do_holes;
- blkmask = (1 << dio->blkfactor) - 1;
- dio_remainder = (dio->block_in_file & blkmask);
+ blkmask = (1 << sdio->blkfactor) - 1;
+ dio_remainder = (sdio->block_in_file & blkmask);
/*
* If we are at the start of IO and that IO
@@ -899,8 +909,8 @@ static int do_direct_IO(struct dio *dio)
* on-disk
*/
if (!buffer_new(map_bh))
- dio->next_block_for_io += dio_remainder;
- dio->blocks_available -= dio_remainder;
+ sdio->next_block_for_io += dio_remainder;
+ sdio->blocks_available -= dio_remainder;
}
do_holes:
/* Handle holes */
@@ -919,7 +929,7 @@ do_holes:
*/
i_size_aligned = ALIGN(i_size_read(dio->inode),
1 << blkbits);
- if (dio->block_in_file >=
+ if (sdio->block_in_file >=
i_size_aligned >> blkbits) {
/* We hit eof */
page_cache_release(page);
@@ -927,7 +937,7 @@ do_holes:
}
zero_user(page, block_in_page << blkbits,
1 << blkbits);
- dio->block_in_file++;
+ sdio->block_in_file++;
block_in_page++;
goto next_block;
}
@@ -937,38 +947,38 @@ do_holes:
* is finer than the underlying fs, go check to see if
* we must zero out the start of this block.
*/
- if (unlikely(dio->blkfactor && !dio->start_zero_done))
- dio_zero_block(dio, 0);
+ if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
+ dio_zero_block(dio, sdio, 0);
/*
* Work out, in this_chunk_blocks, how much disk we
* can add to this page
*/
- this_chunk_blocks = dio->blocks_available;
+ this_chunk_blocks = sdio->blocks_available;
u = (PAGE_SIZE - offset_in_page) >> blkbits;
if (this_chunk_blocks > u)
this_chunk_blocks = u;
- u = dio->final_block_in_request - dio->block_in_file;
+ u = sdio->final_block_in_request - sdio->block_in_file;
if (this_chunk_blocks > u)
this_chunk_blocks = u;
this_chunk_bytes = this_chunk_blocks << blkbits;
BUG_ON(this_chunk_bytes == 0);
- dio->boundary = buffer_boundary(map_bh);
- ret = submit_page_section(dio, page, offset_in_page,
- this_chunk_bytes, dio->next_block_for_io);
+ sdio->boundary = buffer_boundary(map_bh);
+ ret = submit_page_section(dio, sdio, page, offset_in_page,
+ this_chunk_bytes, sdio->next_block_for_io);
if (ret) {
page_cache_release(page);
goto out;
}
- dio->next_block_for_io += this_chunk_blocks;
+ sdio->next_block_for_io += this_chunk_blocks;
- dio->block_in_file += this_chunk_blocks;
+ sdio->block_in_file += this_chunk_blocks;
block_in_page += this_chunk_blocks;
- dio->blocks_available -= this_chunk_blocks;
+ sdio->blocks_available -= this_chunk_blocks;
next_block:
- BUG_ON(dio->block_in_file > dio->final_block_in_request);
- if (dio->block_in_file == dio->final_block_in_request)
+ BUG_ON(sdio->block_in_file > sdio->final_block_in_request);
+ if (sdio->block_in_file == sdio->final_block_in_request)
break;
}
@@ -987,7 +997,7 @@ static ssize_t
direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
const struct iovec *iov, loff_t offset, unsigned long nr_segs,
unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
- dio_submit_t submit_io, struct dio *dio)
+ dio_submit_t submit_io, struct dio *dio, struct dio_submit *sdio)
{
unsigned long user_addr;
unsigned long flags;
@@ -998,15 +1008,15 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
dio->inode = inode;
dio->rw = rw;
- dio->blkbits = blkbits;
- dio->blkfactor = inode->i_blkbits - blkbits;
- dio->block_in_file = offset >> blkbits;
+ sdio->blkbits = blkbits;
+ sdio->blkfactor = inode->i_blkbits - blkbits;
+ sdio->block_in_file = offset >> blkbits;
- dio->get_block = get_block;
+ sdio->get_block = get_block;
dio->end_io = end_io;
- dio->submit_io = submit_io;
- dio->final_block_in_bio = -1;
- dio->next_block_for_io = -1;
+ sdio->submit_io = submit_io;
+ sdio->final_block_in_bio = -1;
+ sdio->next_block_for_io = -1;
dio->iocb = iocb;
dio->i_size = i_size_read(inode);
@@ -1018,45 +1028,45 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
* In case of non-aligned buffers, we may need 2 more
* pages since we need to zero out first and last block.
*/
- if (unlikely(dio->blkfactor))
- dio->pages_in_io = 2;
+ if (unlikely(sdio->blkfactor))
+ sdio->pages_in_io = 2;
for (seg = 0; seg < nr_segs; seg++) {
user_addr = (unsigned long)iov[seg].iov_base;
- dio->pages_in_io +=
+ sdio->pages_in_io +=
((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
- user_addr/PAGE_SIZE);
}
for (seg = 0; seg < nr_segs; seg++) {
user_addr = (unsigned long)iov[seg].iov_base;
- dio->size += bytes = iov[seg].iov_len;
+ sdio->size += bytes = iov[seg].iov_len;
/* Index into the first page of the first block */
- dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
- dio->final_block_in_request = dio->block_in_file +
+ sdio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
+ sdio->final_block_in_request = sdio->block_in_file +
(bytes >> blkbits);
/* Page fetching state */
- dio->head = 0;
- dio->tail = 0;
- dio->curr_page = 0;
+ sdio->head = 0;
+ sdio->tail = 0;
+ sdio->curr_page = 0;
- dio->total_pages = 0;
+ sdio->total_pages = 0;
if (user_addr & (PAGE_SIZE-1)) {
- dio->total_pages++;
+ sdio->total_pages++;
bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
}
- dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
- dio->curr_user_address = user_addr;
+ sdio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
+ sdio->curr_user_address = user_addr;
- ret = do_direct_IO(dio);
+ ret = do_direct_IO(dio, sdio);
dio->result += iov[seg].iov_len -
- ((dio->final_block_in_request - dio->block_in_file) <<
+ ((sdio->final_block_in_request - sdio->block_in_file) <<
blkbits);
if (ret) {
- dio_cleanup(dio);
+ dio_cleanup(dio, sdio);
break;
}
} /* end iovec loop */
@@ -1072,23 +1082,23 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
* There may be some unwritten disk at the end of a part-written
* fs-block-sized block. Go zero that now.
*/
- dio_zero_block(dio, 1);
+ dio_zero_block(dio, sdio, 1);
- if (dio->cur_page) {
- ret2 = dio_send_cur_page(dio);
+ if (sdio->cur_page) {
+ ret2 = dio_send_cur_page(dio, sdio);
if (ret == 0)
ret = ret2;
- page_cache_release(dio->cur_page);
- dio->cur_page = NULL;
+ page_cache_release(sdio->cur_page);
+ sdio->cur_page = NULL;
}
- if (dio->bio)
- dio_bio_submit(dio);
+ if (sdio->bio)
+ dio_bio_submit(dio, sdio);
/*
* It is possible that, we return short IO due to end of file.
* In that case, we need to release all the pages we got hold on.
*/
- dio_cleanup(dio);
+ dio_cleanup(dio, sdio);
/*
* All block lookups have been performed. For READ requests
@@ -1107,7 +1117,7 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
*/
BUG_ON(ret == -EIOCBQUEUED);
if (dio->is_async && ret == 0 && dio->result &&
- ((rw & READ) || (dio->result == dio->size)))
+ ((rw & READ) || (dio->result == sdio->size)))
ret = -EIOCBQUEUED;
if (ret != -EIOCBQUEUED)
@@ -1171,6 +1181,7 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
ssize_t retval = -EINVAL;
loff_t end = offset;
struct dio *dio;
+ struct dio_submit sdio = { 0, };
if (rw & WRITE)
rw = WRITE_ODIRECT;
@@ -1248,7 +1259,7 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
retval = direct_io_worker(rw, iocb, inode, iov, offset,
nr_segs, blkbits, get_block, end_io,
- submit_io, dio);
+ submit_io, dio, &sdio);
out:
return retval;
--
1.7.4.4
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH 2/6] DIO: Fix a wrong comment
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
2011-06-22 19:18 ` [PATCH 1/6] DIO: Separate fields only used in the submission path from struct dio Andi Kleen
@ 2011-06-22 19:18 ` Andi Kleen
2011-06-22 19:18 ` [PATCH 3/6] DIO: Rearrange fields in dio/dio_submit to avoid holes Andi Kleen
` (4 subsequent siblings)
6 siblings, 0 replies; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 19:18 UTC (permalink / raw)
To: linux-fsdevel; +Cc: linux-kernel, dehrenberg, Andi Kleen
From: Andi Kleen <ak@linux.intel.com>
There's nothing on the stack, even before my changes.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
fs/direct-io.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index e1360c6..823d2fe 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -39,7 +39,7 @@
/*
* How many user pages to map in one call to get_user_pages(). This determines
- * the size of a structure on the stack.
+ * the size of a structure in the slab cache
*/
#define DIO_PAGES 64
--
1.7.4.4
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH 3/6] DIO: Rearrange fields in dio/dio_submit to avoid holes.
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
2011-06-22 19:18 ` [PATCH 1/6] DIO: Separate fields only used in the submission path from struct dio Andi Kleen
2011-06-22 19:18 ` [PATCH 2/6] DIO: Fix a wrong comment Andi Kleen
@ 2011-06-22 19:18 ` Andi Kleen
2011-06-22 19:18 ` [PATCH 4/6] DIO: Use a slab cache for struct dio Andi Kleen
` (3 subsequent siblings)
6 siblings, 0 replies; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 19:18 UTC (permalink / raw)
To: linux-fsdevel; +Cc: linux-kernel, dehrenberg, Andi Kleen
From: Andi Kleen <ak@linux.intel.com>
Fix most problems reported by pahole.
There is still a weird 104 byte hole after map_bh. I'm not sure what
causes this.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
fs/direct-io.c | 13 ++++++-------
1 files changed, 6 insertions(+), 7 deletions(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index 823d2fe..03deca3 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -73,10 +73,10 @@ struct dio_submit {
sector_t block_in_file; /* Current offset into the underlying
file in dio_block units. */
unsigned blocks_available; /* At block_in_file. changes */
+ int reap_counter; /* rate limit reaping */
sector_t final_block_in_request;/* doesn't change */
unsigned first_block_in_page; /* doesn't change, Used only once */
int boundary; /* prev block is at a boundary */
- int reap_counter; /* rate limit reaping */
get_block_t *get_block; /* block mapping function */
dio_submit_t *submit_io; /* IO submition function */
@@ -114,27 +114,26 @@ struct dio_submit {
/* dio_state communicated between submission path and end_io */
struct dio {
int flags; /* doesn't change */
- struct inode *inode;
int rw;
+ struct inode *inode;
loff_t i_size; /* i_size when submitted */
dio_iodone_t *end_io; /* IO completion function */
- struct buffer_head map_bh; /* last get_block() result */
/* BIO completion state */
spinlock_t bio_lock; /* protects BIO fields below */
+ int page_errors; /* errno from get_user_pages() */
+ int is_async; /* is IO async ? */
+ int io_error; /* IO error in completion path */
unsigned long refcount; /* direct_io_worker() and bios */
struct bio *bio_list; /* singly linked via bi_private */
struct task_struct *waiter; /* waiting task (NULL if none) */
/* AIO related stuff */
struct kiocb *iocb; /* kiocb */
- int is_async; /* is IO async ? */
- int io_error; /* IO error in completion path */
ssize_t result; /* IO result */
- int page_errors; /* errno from get_user_pages() */
-
+ struct buffer_head map_bh; /* last get_block() result */
/*
* pages[] (and any fields placed after it) are not zeroed out at
* allocation time. Don't add new fields after pages[] unless you
--
1.7.4.4
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH 4/6] DIO: Use a slab cache for struct dio
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
` (2 preceding siblings ...)
2011-06-22 19:18 ` [PATCH 3/6] DIO: Rearrange fields in dio/dio_submit to avoid holes Andi Kleen
@ 2011-06-22 19:18 ` Andi Kleen
2011-06-22 19:18 ` [PATCH 5/6] DIO: Separate map_bh from dio Andi Kleen
` (2 subsequent siblings)
6 siblings, 0 replies; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 19:18 UTC (permalink / raw)
To: linux-fsdevel; +Cc: linux-kernel, dehrenberg, Andi Kleen
From: Andi Kleen <ak@linux.intel.com>
A direct slab call is slightly faster than kmalloc and can be better cached
per CPU. It also avoids rounding to the next kmalloc slab.
In addition this enforces cache line alignment for struct dio to avoid
any false sharing.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
fs/direct-io.c | 19 ++++++++++++++-----
1 files changed, 14 insertions(+), 5 deletions(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index 03deca3..49dbd06 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -140,7 +140,9 @@ struct dio {
* wish that they not be zeroed.
*/
struct page *pages[DIO_PAGES]; /* page buffer */
-};
+} ____cacheline_aligned_in_smp;
+
+static struct kmem_cache *dio_cache __read_mostly;
/*
* How many pages are in the queue?
@@ -288,7 +290,7 @@ static void dio_bio_end_aio(struct bio *bio, int error)
if (remaining == 0) {
dio_complete(dio, dio->iocb->ki_pos, 0, true);
- kfree(dio);
+ kmem_cache_free(dio_cache, dio);
}
}
@@ -1139,7 +1141,7 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
if (ret2 == 0) {
ret = dio_complete(dio, offset, ret, false);
- kfree(dio);
+ kmem_cache_free(dio_cache, dio);
} else
BUG_ON(ret != -EIOCBQUEUED);
@@ -1210,7 +1212,7 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
}
}
- dio = kmalloc(sizeof(*dio), GFP_KERNEL);
+ dio = kmem_cache_alloc(dio_cache, GFP_KERNEL);
retval = -ENOMEM;
if (!dio)
goto out;
@@ -1235,7 +1237,7 @@ __blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
end - 1);
if (retval) {
mutex_unlock(&inode->i_mutex);
- kfree(dio);
+ kmem_cache_free(dio_cache, dio);
goto out;
}
}
@@ -1264,3 +1266,10 @@ out:
return retval;
}
EXPORT_SYMBOL(__blockdev_direct_IO);
+
+static __init int dio_init(void)
+{
+ dio_cache = KMEM_CACHE(dio, SLAB_PANIC);
+ return 0;
+}
+module_init(dio_init)
--
1.7.4.4
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH 5/6] DIO: Separate map_bh from dio
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
` (3 preceding siblings ...)
2011-06-22 19:18 ` [PATCH 4/6] DIO: Use a slab cache for struct dio Andi Kleen
@ 2011-06-22 19:18 ` Andi Kleen
2011-06-22 19:59 ` Christoph Hellwig
2011-06-22 19:18 ` [PATCH 6/6] DIO: Inline the complete submission path Andi Kleen
2011-07-27 21:06 ` RFC: Micro-optimize direct IO " Christoph Hellwig
6 siblings, 1 reply; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 19:18 UTC (permalink / raw)
To: linux-fsdevel; +Cc: linux-kernel, dehrenberg, Andi Kleen
From: Andi Kleen <ak@linux.intel.com>
Only a single b_private field in the map_bh buffer head is needed after
the submission path. Move map_bh separately to avoid storing
this information in the long term slab.
This avoids the weird 104 byte hole in struct dio_submit which also needed
to be memseted early.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
fs/direct-io.c | 65 +++++++++++++++++++++++++++++++-------------------------
1 files changed, 36 insertions(+), 29 deletions(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index 49dbd06..cc75445 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -119,6 +119,7 @@ struct dio {
loff_t i_size; /* i_size when submitted */
dio_iodone_t *end_io; /* IO completion function */
+ void *b_private; /* copy from map_bh.b_private */
/* BIO completion state */
spinlock_t bio_lock; /* protects BIO fields below */
@@ -133,7 +134,6 @@ struct dio {
struct kiocb *iocb; /* kiocb */
ssize_t result; /* IO result */
- struct buffer_head map_bh; /* last get_block() result */
/*
* pages[] (and any fields placed after it) are not zeroed out at
* allocation time. Don't add new fields after pages[] unless you
@@ -257,7 +257,7 @@ static ssize_t dio_complete(struct dio *dio, loff_t offset, ssize_t ret, bool is
if (dio->end_io && dio->result) {
dio->end_io(dio->iocb, offset, transferred,
- dio->map_bh.b_private, ret, is_async);
+ dio->b_private, ret, is_async);
} else if (is_async) {
aio_complete(dio->iocb, ret, 0);
}
@@ -532,10 +532,10 @@ static int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
* buffer_mapped(). However the direct-io code will only process holes one
* block at a time - it will repeatedly call get_block() as it walks the hole.
*/
-static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
+static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
+ struct buffer_head *map_bh)
{
int ret;
- struct buffer_head *map_bh = &dio->map_bh;
sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
unsigned long fs_count; /* Number of filesystem-sized blocks */
unsigned long dio_count;/* Number of dio_block-sized blocks */
@@ -579,6 +579,9 @@ static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
ret = (*sdio->get_block)(dio->inode, fs_startblk,
map_bh, create);
+
+ /* Store for completion */
+ dio->b_private = map_bh->b_private;
}
return ret;
}
@@ -587,7 +590,7 @@ static int get_more_blocks(struct dio *dio, struct dio_submit *sdio)
* There is no bio. Make one now.
*/
static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
- sector_t start_sector)
+ sector_t start_sector, struct buffer_head *map_bh)
{
sector_t sector;
int ret, nr_pages;
@@ -596,10 +599,10 @@ static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
if (ret)
goto out;
sector = start_sector << (sdio->blkbits - 9);
- nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
+ nr_pages = min(sdio->pages_in_io, bio_get_nr_vecs(map_bh->b_bdev));
nr_pages = min(nr_pages, BIO_MAX_PAGES);
BUG_ON(nr_pages <= 0);
- dio_bio_alloc(dio, sdio, dio->map_bh.b_bdev, sector, nr_pages);
+ dio_bio_alloc(dio, sdio, map_bh->b_bdev, sector, nr_pages);
sdio->boundary = 0;
out:
return ret;
@@ -644,7 +647,8 @@ static int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
* The caller of this function is responsible for removing cur_page from the
* dio, and for dropping the refcount which came from that presence.
*/
-static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio)
+static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
+ struct buffer_head *map_bh)
{
int ret = 0;
@@ -679,14 +683,14 @@ static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio)
}
if (sdio->bio == NULL) {
- ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
+ ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
if (ret)
goto out;
}
if (dio_bio_add_page(dio, sdio) != 0) {
dio_bio_submit(dio, sdio);
- ret = dio_new_bio(dio, sdio, sdio->cur_page_block);
+ ret = dio_new_bio(dio, sdio, sdio->cur_page_block, map_bh);
if (ret == 0) {
ret = dio_bio_add_page(dio, sdio);
BUG_ON(ret != 0);
@@ -715,7 +719,8 @@ out:
*/
static int
submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
- unsigned offset, unsigned len, sector_t blocknr)
+ unsigned offset, unsigned len, sector_t blocknr,
+ struct buffer_head *map_bh)
{
int ret = 0;
@@ -740,7 +745,7 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
* avoid metadata seeks.
*/
if (sdio->boundary) {
- ret = dio_send_cur_page(dio, sdio);
+ ret = dio_send_cur_page(dio, sdio, map_bh);
page_cache_release(sdio->cur_page);
sdio->cur_page = NULL;
}
@@ -751,7 +756,7 @@ submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
* If there's a deferred page already there then send it.
*/
if (sdio->cur_page) {
- ret = dio_send_cur_page(dio, sdio);
+ ret = dio_send_cur_page(dio, sdio, map_bh);
page_cache_release(sdio->cur_page);
sdio->cur_page = NULL;
if (ret)
@@ -773,16 +778,16 @@ out:
* file blocks. Only called for S_ISREG files - blockdevs do not set
* buffer_new
*/
-static void clean_blockdev_aliases(struct dio *dio)
+static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
{
unsigned i;
unsigned nblocks;
- nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
+ nblocks = map_bh->b_size >> dio->inode->i_blkbits;
for (i = 0; i < nblocks; i++) {
- unmap_underlying_metadata(dio->map_bh.b_bdev,
- dio->map_bh.b_blocknr + i);
+ unmap_underlying_metadata(map_bh->b_bdev,
+ map_bh->b_blocknr + i);
}
}
@@ -795,7 +800,8 @@ static void clean_blockdev_aliases(struct dio *dio)
* `end' is zero if we're doing the start of the IO, 1 at the end of the
* IO.
*/
-static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
+static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end,
+ struct buffer_head *map_bh)
{
unsigned dio_blocks_per_fs_block;
unsigned this_chunk_blocks; /* In dio_blocks */
@@ -803,7 +809,7 @@ static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
struct page *page;
sdio->start_zero_done = 1;
- if (!sdio->blkfactor || !buffer_new(&dio->map_bh))
+ if (!sdio->blkfactor || !buffer_new(map_bh))
return;
dio_blocks_per_fs_block = 1 << sdio->blkfactor;
@@ -823,7 +829,7 @@ static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
page = ZERO_PAGE(0);
if (submit_page_section(dio, sdio, page, 0, this_chunk_bytes,
- sdio->next_block_for_io))
+ sdio->next_block_for_io, map_bh))
return;
sdio->next_block_for_io += this_chunk_blocks;
@@ -845,13 +851,13 @@ static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end)
* it should set b_size to PAGE_SIZE or more inside get_block(). This gives
* fine alignment but still allows this function to work in PAGE_SIZE units.
*/
-static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
+static int do_direct_IO(struct dio *dio, struct dio_submit *sdio,
+ struct buffer_head *map_bh)
{
const unsigned blkbits = sdio->blkbits;
const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
struct page *page;
unsigned block_in_page;
- struct buffer_head *map_bh = &dio->map_bh;
int ret = 0;
/* The I/O can start at any block offset within the first page */
@@ -877,7 +883,7 @@ static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
unsigned long blkmask;
unsigned long dio_remainder;
- ret = get_more_blocks(dio, sdio);
+ ret = get_more_blocks(dio, sdio, map_bh);
if (ret) {
page_cache_release(page);
goto out;
@@ -890,7 +896,7 @@ static int do_direct_IO(struct dio *dio, struct dio_submit *sdio)
sdio->next_block_for_io =
map_bh->b_blocknr << sdio->blkfactor;
if (buffer_new(map_bh))
- clean_blockdev_aliases(dio);
+ clean_blockdev_aliases(dio, map_bh);
if (!sdio->blkfactor)
goto do_holes;
@@ -949,7 +955,7 @@ do_holes:
* we must zero out the start of this block.
*/
if (unlikely(sdio->blkfactor && !sdio->start_zero_done))
- dio_zero_block(dio, sdio, 0);
+ dio_zero_block(dio, sdio, 0, map_bh);
/*
* Work out, in this_chunk_blocks, how much disk we
@@ -967,7 +973,7 @@ do_holes:
sdio->boundary = buffer_boundary(map_bh);
ret = submit_page_section(dio, sdio, page, offset_in_page,
- this_chunk_bytes, sdio->next_block_for_io);
+ this_chunk_bytes, sdio->next_block_for_io, map_bh);
if (ret) {
page_cache_release(page);
goto out;
@@ -1006,6 +1012,7 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
ssize_t ret = 0;
ssize_t ret2;
size_t bytes;
+ struct buffer_head map_bh = { 0, };
dio->inode = inode;
dio->rw = rw;
@@ -1060,7 +1067,7 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
sdio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
sdio->curr_user_address = user_addr;
- ret = do_direct_IO(dio, sdio);
+ ret = do_direct_IO(dio, sdio, &map_bh);
dio->result += iov[seg].iov_len -
((sdio->final_block_in_request - sdio->block_in_file) <<
@@ -1083,10 +1090,10 @@ direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
* There may be some unwritten disk at the end of a part-written
* fs-block-sized block. Go zero that now.
*/
- dio_zero_block(dio, sdio, 1);
+ dio_zero_block(dio, sdio, 1, &map_bh);
if (sdio->cur_page) {
- ret2 = dio_send_cur_page(dio, sdio);
+ ret2 = dio_send_cur_page(dio, sdio, &map_bh);
if (ret == 0)
ret = ret2;
page_cache_release(sdio->cur_page);
--
1.7.4.4
^ permalink raw reply related [flat|nested] 13+ messages in thread* [PATCH 6/6] DIO: Inline the complete submission path
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
` (4 preceding siblings ...)
2011-06-22 19:18 ` [PATCH 5/6] DIO: Separate map_bh from dio Andi Kleen
@ 2011-06-22 19:18 ` Andi Kleen
2011-06-22 19:59 ` Christoph Hellwig
2011-07-27 21:06 ` RFC: Micro-optimize direct IO " Christoph Hellwig
6 siblings, 1 reply; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 19:18 UTC (permalink / raw)
To: linux-fsdevel; +Cc: linux-kernel, dehrenberg, Andi Kleen
From: Andi Kleen <ak@linux.intel.com>
Add inlines to all the submission path functions. While this increases
code slightly it also gives gcc a lot of optimization opportunities
in this critical hotpath.
In particular -- together with some other changes -- this
allows gcc to get rid of the unnecessary clearing of
sdio at the beginning and optimize the messy parameter passing.
Note that benefits are only seen with CONFIG_OPTIMIZE_INLINING
and CONFIG_CC_OPTIMIZE_FOR_SIZE both set to off.
Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
fs/direct-io.c | 26 +++++++++++++-------------
1 files changed, 13 insertions(+), 13 deletions(-)
diff --git a/fs/direct-io.c b/fs/direct-io.c
index cc75445..1a2edbf 100644
--- a/fs/direct-io.c
+++ b/fs/direct-io.c
@@ -155,7 +155,7 @@ static inline unsigned dio_pages_present(struct dio *dio, struct dio_submit *sdi
/*
* Go grab and pin some userspace pages. Typically we'll get 64 at a time.
*/
-static int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
+static inline int dio_refill_pages(struct dio *dio, struct dio_submit *sdio)
{
int ret;
int nr_pages;
@@ -201,7 +201,7 @@ out:
* decent number of pages, less frequently. To provide nicer use of the
* L1 cache.
*/
-static struct page *dio_get_page(struct dio *dio, struct dio_submit *sdio)
+static inline struct page *dio_get_page(struct dio *dio, struct dio_submit *sdio)
{
if (dio_pages_present(dio, sdio) == 0) {
int ret;
@@ -334,7 +334,7 @@ void dio_end_io(struct bio *bio, int error)
}
EXPORT_SYMBOL_GPL(dio_end_io);
-static void
+static void inline
dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
struct block_device *bdev,
sector_t first_sector, int nr_vecs)
@@ -365,7 +365,7 @@ dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,
*
* bios hold a dio reference between submit_bio and ->end_io.
*/
-static void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
+static inline void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
{
struct bio *bio = sdio->bio;
unsigned long flags;
@@ -393,7 +393,7 @@ static void dio_bio_submit(struct dio *dio, struct dio_submit *sdio)
/*
* Release any resources in case of a failure
*/
-static void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
+static inline void dio_cleanup(struct dio *dio, struct dio_submit *sdio)
{
while (dio_pages_present(dio, sdio))
page_cache_release(dio_get_page(dio, sdio));
@@ -486,7 +486,7 @@ static void dio_await_completion(struct dio *dio)
*
* This also helps to limit the peak amount of pinned userspace memory.
*/
-static int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
+static inline int dio_bio_reap(struct dio *dio, struct dio_submit *sdio)
{
int ret = 0;
@@ -589,7 +589,7 @@ static int get_more_blocks(struct dio *dio, struct dio_submit *sdio,
/*
* There is no bio. Make one now.
*/
-static int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
+static inline int dio_new_bio(struct dio *dio, struct dio_submit *sdio,
sector_t start_sector, struct buffer_head *map_bh)
{
sector_t sector;
@@ -615,7 +615,7 @@ out:
*
* Return zero on success. Non-zero means the caller needs to start a new BIO.
*/
-static int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
+static inline int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
{
int ret;
@@ -647,8 +647,8 @@ static int dio_bio_add_page(struct dio *dio, struct dio_submit *sdio)
* The caller of this function is responsible for removing cur_page from the
* dio, and for dropping the refcount which came from that presence.
*/
-static int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
- struct buffer_head *map_bh)
+static inline int dio_send_cur_page(struct dio *dio, struct dio_submit *sdio,
+ struct buffer_head *map_bh)
{
int ret = 0;
@@ -717,7 +717,7 @@ out:
* If that doesn't work out then we put the old page into the bio and add this
* page to the dio instead.
*/
-static int
+static inline int
submit_page_section(struct dio *dio, struct dio_submit *sdio, struct page *page,
unsigned offset, unsigned len, sector_t blocknr,
struct buffer_head *map_bh)
@@ -800,7 +800,7 @@ static void clean_blockdev_aliases(struct dio *dio, struct buffer_head *map_bh)
* `end' is zero if we're doing the start of the IO, 1 at the end of the
* IO.
*/
-static void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end,
+static inline void dio_zero_block(struct dio *dio, struct dio_submit *sdio, int end,
struct buffer_head *map_bh)
{
unsigned dio_blocks_per_fs_block;
@@ -1000,7 +1000,7 @@ out:
/*
* Releases both i_mutex and i_alloc_sem
*/
-static ssize_t
+static inline ssize_t
direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
const struct iovec *iov, loff_t offset, unsigned long nr_segs,
unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
--
1.7.4.4
^ permalink raw reply related [flat|nested] 13+ messages in thread* Re: [PATCH 6/6] DIO: Inline the complete submission path
2011-06-22 19:18 ` [PATCH 6/6] DIO: Inline the complete submission path Andi Kleen
@ 2011-06-22 19:59 ` Christoph Hellwig
2011-06-22 20:10 ` Andi Kleen
0 siblings, 1 reply; 13+ messages in thread
From: Christoph Hellwig @ 2011-06-22 19:59 UTC (permalink / raw)
To: Andi Kleen; +Cc: linux-fsdevel, linux-kernel, dehrenberg, Andi Kleen
> -static ssize_t
> +static inline ssize_t
> direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
> const struct iovec *iov, loff_t offset, unsigned long nr_segs,
> unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
For this one please just merge it into the caller. There's no logical
separation between the two anyway.
^ permalink raw reply [flat|nested] 13+ messages in thread
* Re: [PATCH 6/6] DIO: Inline the complete submission path
2011-06-22 19:59 ` Christoph Hellwig
@ 2011-06-22 20:10 ` Andi Kleen
2011-06-23 10:35 ` Christoph Hellwig
0 siblings, 1 reply; 13+ messages in thread
From: Andi Kleen @ 2011-06-22 20:10 UTC (permalink / raw)
To: Christoph Hellwig
Cc: Andi Kleen, linux-fsdevel, linux-kernel, dehrenberg, Andi Kleen
On Wed, Jun 22, 2011 at 03:59:56PM -0400, Christoph Hellwig wrote:
> > -static ssize_t
> > +static inline ssize_t
> > direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
> > const struct iovec *iov, loff_t offset, unsigned long nr_segs,
> > unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
>
> For this one please just merge it into the caller. There's no logical
> separation between the two anyway.
The compiler does that anyways.
Or did you mean in source? That would conflict with the coding standard.
-Andi
--
ak@linux.intel.com -- Speaking for myself only.
^ permalink raw reply [flat|nested] 13+ messages in thread
* Re: RFC: Micro-optimize direct IO submission path
2011-06-22 19:18 RFC: Micro-optimize direct IO submission path Andi Kleen
` (5 preceding siblings ...)
2011-06-22 19:18 ` [PATCH 6/6] DIO: Inline the complete submission path Andi Kleen
@ 2011-07-27 21:06 ` Christoph Hellwig
2011-07-27 21:35 ` Andi Kleen
6 siblings, 1 reply; 13+ messages in thread
From: Christoph Hellwig @ 2011-07-27 21:06 UTC (permalink / raw)
To: Andi Kleen; +Cc: linux-fsdevel, linux-kernel, dehrenberg
Any chance you could resend these against current mainline? It would
be sad to miss these easy optimizations.
^ permalink raw reply [flat|nested] 13+ messages in thread* Re: RFC: Micro-optimize direct IO submission path
2011-07-27 21:06 ` RFC: Micro-optimize direct IO " Christoph Hellwig
@ 2011-07-27 21:35 ` Andi Kleen
0 siblings, 0 replies; 13+ messages in thread
From: Andi Kleen @ 2011-07-27 21:35 UTC (permalink / raw)
To: Christoph Hellwig; +Cc: Andi Kleen, linux-fsdevel, linux-kernel, dehrenberg
On Wed, Jul 27, 2011 at 05:06:06PM -0400, Christoph Hellwig wrote:
> Any chance you could resend these against current mainline? It would
> be sad to miss these easy optimizations.
I was waiting for the benchmarkers to rerun them.
I had done an additional prefetch optimization that seemed to give
1% more on a big benchmark for one result, but there was another
suspicious one.
Hopefully soon.
-Andi
--
ak@linux.intel.com -- Speaking for myself only.
^ permalink raw reply [flat|nested] 13+ messages in thread